DeepSeek launched its R1 model on January 20, 2025, and instantly became one of the most talked about AI products in the world. A Chinese AI startup founded in July 2023, it built a reasoning model that matches or outperforms OpenAI on several benchmarks at a fraction of the cost. The release wiped $600 billion from Nvidia’s market cap in a single day and forced every major AI lab to revisit its pricing strategy.
This article covers verified DeepSeek statistics across 14 dimensions: users, downloads, geography, demographics, training economics, benchmarks, model specs, API pricing, and more. Let’s deep dive and find out the latest DeepSeek stats.
Key DeepSeek Statistics at a Glance
96.88 million monthly active users as of April 2025
130 million monthly active users by the end of 2025
22.15 million peak daily users as of January 2025
50 million and a 4.1 rating on Google Play from 0.293 million reviews as of June 2026
175 million estimated total downloads by the end of 2025
DeepSeek V3 cost $5.576 million to train, roughly 1/18th the estimated cost of GPT-4
NVIDIA lost approximately $600 billion in market cap on January 27, 2025
DeepSeek API is up to 35x cheaper than Claude Sonnet per million tokens
Potential $45 billion valuation as of May 2026, with Tencent and Alibaba in talks
DeepSeek Users Statistics
DeepSeek’s user growth after the R1 launch was one of the fastest adaptation curves recorded for any AI product:
| Period | Monthly Active Users |
|---|---|
| January 2025 | 33.7 million |
| March 2025 | 77 million |
| April 2025 | 96.88 million |
| End of 2025 | 130 million |
| March 2026 (Website visits) | 350.8 million |
Source: Aicpb.com| Business of Apps
The April 2025 figure represents a 25.81% month-over-month increase from March 2025, placing DeepSeek at #4 among global AI apps by active user base.
Daily Active Users and Website Visitors
Weekly website visitors: 15.9 million: Backlinko
Daily visitors grew from 7,475 (August 2024) to 22.15 million by January 2025, a 312% spike coinciding with the R1 release: DemandSage
App Download Statistics
DeepSeek became the #1 most downloaded app in the App Store across 156 countries within days of the R1 launch.
Total Downloads
57.2 million+ downloads by May 22, 2025 (Google Play + App Store combined): Appfigures
34.6 million downloads from Google Play by May 2025: SQ Magazine
22.6 million downloads from the App Store by May 2025: SQ Magazine
Estimated 175 million total downloads by the end of 2025: Business of Apps
Monthly Download Breakdown (2025)
| Period | Downloads | Source |
|---|---|---|
| January 2025 | 14.2 million | [Backlinko](https://backlinko.com/deepseek-stats) |
| February 2025 | 19.6 million | [Backlinko](https://backlinko.com/deepseek-stats) |
| March 2025 | 9 million | [Backlinko](https://backlinko.com/deepseek-stats) |
| April 2025 | 9 million | [Backlinko](https://backlinko.com/deepseek-stats) |
| May 2025 (partial, as of May 22) | 5.4 million | [Backlinko](https://backlinko.com/deepseek-stats) |
| Q1 2025 total | ~83 million | [Business of Apps](https://www.businessofapps.com/data/deepseek-statistics/) |
| Q2 2025 | ~39 million | [Business of Apps](https://www.businessofapps.com/data/deepseek-statistics/) |
| Q3 2025 | ~20 million | [Business of Apps](https://www.businessofapps.com/data/deepseek-statistics/) |
| Q4 2025 | ~18 million | [Business of Apps](https://www.businessofapps.com/data/deepseek-statistics/) |
| Q1 2026 | ~13 million | [Business of Apps](https://www.businessofapps.com/data/deepseek-statistics/) |
Note:* Appfigures covers 99 countries; Business of Apps covers a broader dataset. Quarterly figures may differ due to methodology.*
Downloads by Country
| Country | Share of Downloads |
|---|---|
| China | 34% |
| India | 8% |
| Russia | 7% |
| United States | 6% |
| Pakistan | 4% |
| Brazil | 4% |
| Indonesia | 4% |
| France | 3% |
| UK | 3% |
| Other Countries | 27% |
Source: Business of Apps
China accounts for the largest single share at 34%, driven by domestic access through the DeepSeek web interface and app. India and Russia rank second and third, reflecting strong demand in emerging AI markets.
User Demographics by Platform
DeepSeek skews young across both platforms, with 18- 24-year-olds representing the largest age cohort.
| Age Group | iOS Users | Android Users |
|---|---|---|
| 18-24 | 38.7% | 44.9% |
| 25-34 | 22.1% | 13.2% |
| 35-49 | 15.3% | 14.9% |
| 50-64 | 23.3% | 26.1% |
| 65+ | 0.6% | 1.0% |
Source: Backlinko | DemandSage
Android users aged 18-24 (44.9%) outpace iOS users in the same bracket (38.7%)
The 50-64 group accounts for a notable 23-26% across both platforms, suggesting professional and academic adoption alongside Gen Z
Both iOS and Android user bases lean male
Training Cost Economics
The DeepSeek-V3 training cost story is the single most cited data point in the AI industry in 2025. It reframed how investors and engineers think about frontier model economics.
| Metric | DeepSeek-V3 | GPT-4 (Estimated) |
|---|---|---|
| Training cost | $5.576 million | ~$100 million |
| GPU hours | 2.788 million H800 hours | Not disclosed |
| Cost ratio | 1x | ~18x more expensive |
Source: Arxiv DeepSeek-R1 paper (2501.12948) | DemandSage
DeepSeek used approximately 2,000 Nvidia H800 chips, which are lower-spec export-controlled chips: DemandSage
Training tokens: 14.8 trillion: Arxiv (2501.12948)
Context window: 1million tokens; max output: 384K tokens per response: DocsBot
Model Technical Specifications
DeepSeek uses a Mixture-of-Experts (MoE) architecture, which activates only a subset of parameters per inference call. This is what enables competitive performance at low compute cost.
| Specification | Value |
|---|---|
| Model | DeepSeek-V3 |
| Total parameters | 671 billion |
| Active parameters per token | 37 billion |
| Architecture | Mixture-of-Experts (MoE) |
| Training tokens | 14.8 trillion |
| Context window | 128,000 tokens |
| Max output tokens | 8,000 tokens |
| Open source | Yes (selected versions) |
Source: Arxiv (2501.12948) | DocsBot | DemandSage
The MoE architecture means only 37B out of 671B parameters are active at any given inference step. This drastically reduces compute cost per query compared to a dense model of equivalent total size.
Benchmark Performance: DeepSeek R1 vs. OpenAI o1
DeepSeek-R1 was benchmarked against OpenAI-o1-1217 across reasoning, math, coding, and science tasks.
Reasoning and Math Benchmarks
| Benchmark | DeepSeek-R1 | OpenAI o1-1217 | Winner |
|---|---|---|---|
| MMLU | 90.8% | 91.8% | OpenAI |
| MATH-500 | 97.3% | 96.4% | DeepSeek |
| AIME 2024 | 79.8% | 79.2% | DeepSeek |
| GPQA Diamond | 71.5% | 75.7% | OpenAI |
Coding Benchmarks
| Benchmark | DeepSeek-R1 | OpenAI o1-1217 | Winner |
|---|---|---|---|
| SWE-bench Verified | 49.2% | 48.9% | DeepSeek |
| LiveCodeBench | 65.9% | 63.4% | DeepSeek |
| Codeforces Rating | 2029 | 2061 | OpenAI |
| Codeforces Percentile | 96.3% | 96.6% | OpenAI |
| Aider-Polyglot | 53.3% | 61.7% | OpenAI |
Source: Arxiv (2501.12948) | DemandSage
DeepSeek-R1 wins 2 of 5 coding benchmarks and 2 of 4 math/reasoning benchmarks
DeepSeek leads in general coding (LiveCodeBench, SWE-bench), while OpenAI leads in multilingual coding (Aider-Polyglot)
On MATH-500 and AIME 2024, DeepSeek outperforms OpenAI, showing stronger raw mathematical reasoning
API Pricing Comparison
DeepSeek's API pricing is among the lowest available for frontier-class models.
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| DeepSeek-V3 | $0.14 | $0.28 |
| DeepSeek-R1 | $0.55 | $2.19 |
| DeepSeek V4 Pro | $1.74 | $3.48 |
| GPT-4o | $2.50 | $10.00 |
| Claude Sonnet 4.6 | $3.00 | $15.00 |
Source:** **DocsBot DeepSeek-V4 model card | DemandSage
DeepSeek V3 is approximately 35x cheaper than Claude Sonnet on output tokens: DocsBot
DeepSeek R1 is roughly 5x cheaper than GPT-4o on output
For developers running millions of daily queries, the cost gap is a primary driver of DeepSeek adoption
Valuation and Funding
DeepSeek operated self-funded until 2026, backed by High-Flyer, a Chinese quantitative hedge fund co-founded by CEO Liang Wenfeng.
| Detail | Data |
|---|---|
| Founded | July 2023, Hangzhou, China |
| Employees | 150-200 |
| Primary funder (to April 2026) | High-Flyer Capital |
| Founder stake | ~90% (Liang Wenfeng) |
| First external round (2026) | Targeting a $10 billion valuation |
| Potential valuation (May 2026) | $45 billion |
| Investors in talks | Tencent and Alibaba |
Source: DemandSage | Business of Apps
Liang Wenfeng's background is in quantitative trading, not traditional AI research
DeepSeek hires primarily from top Chinese universities, focusing on fresh graduates
The company operates with a headcount of 150-200, compared to OpenAI's 3,500+
***Note: ***The $45 billion valuation figure was reported in May 2026 and has not been confirmed by DeepSeek directly.
Model Version Timeline
| Model | Release Date |
|---|---|
| DeepSeek Coder V1 | November 2023 |
| DeepSeek LLM 67B | November 2023 |
| DeepSeek-V2 | May 2024 |
| DeepSeek-V2.5 | September 2024 |
| DeepSeek-V3 | December 2024 |
| DeepSeek-R1 | January 20, 2025 |
| DeepSeek-V3.1 | August 2025 |
| DeepSeek V4 Pro / Flash | April 2026 |
Source: Business of Apps
The R1 release on January 20, 2025, was the inflection point for global adoption. V4 Pro and Flash, released in April 2026, introduced updated pricing tiers and improved multilingual performance.
Security and Jailbreak Concerns
DeepSeek's safety profile has become a significant factor for enterprise adoption decisions.
| Model Category | Jailbreak Compliance Rate |
|---|---|
| DeepSeek | 94% |
| US frontier models (average) | 8% |
Source: NIST CAISI evaluation
The NIST CAISI evaluation tested models against a standardized set of adversarial prompts
A 94% jailbreak compliance rate means DeepSeek followed harmful instructions 94% of the time when bypasses were applied
Major hyperscalers (AWS, Azure, Google Cloud) add policy overlays when serving DeepSeek via managed APIs
AWS Bedrock applies additional content filtering on top of the base model: AWS Bedrock
For enterprise teams evaluating DeepSeek, the NIST findings are the primary red flag. Self-hosted deployments carry higher risk without the policy layers added by cloud providers.
Regulatory Bans and Restrictions
Multiple governments and agencies restricted DeepSeek within weeks of the R1 launch, citing data privacy and national security concerns.
| Country / Organization | Restriction Type | Date |
|---|---|---|
| Italy | Removed from app stores | January 30, 2025 |
| Taiwan | Government agency ban | February 3, 2025 |
| Australia | Government agency ban | February 4, 2025 |
| South Korea | Key ministry restriction | February 5, 2025 |
| US Navy | Usage ban | January 2025 |
| NASA | Usage ban | January 31, 2025 |
| Texas (AG investigation) | Initiated investigation | February 2025 |
| US Congress | Internal usage ban | 2025 |
| US Pentagon | Usage restrictions | 2025 |
Source: DemandSage | Business of Apps
South Korea later lifted its ministerial ban after conducting an internal review
Italy's removal followed a formal privacy probe under GDPR; DeepSeek did not comply with regulators' data requests within the deadline
The US bans are agency-level, not a national prohibition
Market Impact: The DeepSeek Shock
The R1 release on January 20, 2025, triggered the largest single-day market cap loss ever caused by a software announcement.
| Metric | Data |
|---|---|
| Nvidia market cap lost (Jan 27, 2025) | ~$600 billion |
| Nvidia stock drop | ~17% |
| Cause | DeepSeek R1 release demonstrated comparable AI at ~1/18th training cost |
| Investor concern | Lower compute demand reduces the need for high-end AI chips |
Source: DemandSage | Business of Apps
The event was widely termed the "DeepSeek Shock" in financial media
Nvidia's single-day loss exceeded the GDP of several mid-sized economies
The drop raised broader questions about the AI infrastructure investment thesis: if frontier models can be trained cheaply, demand for expensive GPU clusters may not grow as projected
Chip stocks, including Broadcom, ASML, and TSMC, also declined on the same day
Enterprise Cloud Adoption
DeepSeek's open-weight models accelerated its adoption inside enterprise cloud environments, where teams can deploy the model without sending data to DeepSeek's servers.
AWS Bedrock made DeepSeek-R1 generally available in March 2025, becoming the first major cloud provider to do so: AWS Bedrock
Thousands of AWS customers deployed DeepSeek-R1 via Bedrock within the first months of availability
Azure AI Foundry added DeepSeek-R1 to its model catalog in February 2025: Azure AI Foundry
Google Cloud Vertex AI listed DeepSeek-R1 as a deployable model in early 2025
Enterprise deployments typically layer additional content filters over the base model, addressing the jailbreak compliance risk noted in the NIST evaluation
Conclusion
DeepSeek went from a little-known Chinese AI lab to one of the most downloaded apps in the world in under six months. Its R1 model combined benchmark-competitive performance with a training cost that exposed a fundamental pricing gap across the AI industry.
The core numbers: 130 million MAUs by the end of 2025, 175 million total downloads, a $5.576M training cost versus GPT-4's estimated $100M, API pricing 35x below Claude Sonnet, and a NIST jailbreak compliance rate that has slowed enterprise adoption without policy overlays.
For developers, the API pricing story is the most actionable takeaway. For investors and enterprises, the security profile and regulatory picture require careful evaluation before production deployment.